Machine learning for early dengue diagnosis using routine clinical and haematological parameters: evidence from a prospective outbreak cohort in Munshiganj, Bangladesh

Dengue fever remains a major public health burden in Bangladesh, which recorded its worst outbreak on record in 2023 with 321,179 cases and 1,705 deaths. This study applied six supervised machine learning algorithms (Logistic Regression, K-Nearest Neighbors, Support Vector Machine, Random Forest, XGBoost, and LightGBM) to a publicly available dataset of 1,018 patients from Life Aid General Hospital and Diagnostic, Munshiganj, Dhaka, Bangladesh, comprising 697 dengue-positive and 321 dengue-negative cases. Features included platelet count, white blood cell count, six clinical symptoms, and demographic variables. Class imbalance was addressed using SMOTE on the training set only. Random Forest and XGBoost both achieved 99.02 percent accuracy and AUC-ROC of 0.9950, confirmed by 10-fold cross-validation at 98.72 percent (SD 0.63 percent). A CBC-only analysis using only platelet count and WBC count achieved identical accuracy with XGBoost and LightGBM. SHAP analysis identified platelet count (mean absolute SHAP 3.657) and WBC count (3.656) as the dominant predictors by a large margin. These findings support the integration of machine learning-based clinical decision support tools into dengue triage workflows in resource-limited settings in Bangladesh and across South Asia.